from: "Edward R. Cook" <drdendroatXYZxyzo.columbia.edu>
subject: Extended PDSI recons
to: Keith Briffa <k.briffaatXYZxyz.ac.uk>
I have put the file of extended PDSI recons on my ftp site. So, go to
ftp.ldeo.columbia.edu and log on anonymously in the usual way:
User id: anonymous
Password: your email address
Then, cd to /pub/cook/briffa and there you will find the file called
PDSI_EXT.TAB. The grid point data are stored columnwise, one column per
grid point. The grid point number is at the top of each column. Now, in
reality the column data are NOT tab delimited. Rather, they are space
delimited with decimal points (.) used for missing values (the SAS
convention I believe). The file should be easy to read into any kind of
spread sheet. These data ought not to be considered final products, as I
may modify the way in which I extended the series back in time using PC
regression. However, they are not likely to change all that much either.
The big adjustment I made to the extended series was interms of variance.
Because each model extension produced a somewhat different R-square, the
variance would normally change with R-square. So, I artificially adjusted
the variance to take into account and remove the effects of time-varying
R-square on the estimates.
I am not totally happy with adjusting the variance as I did. However, it is
probably no worse then allowing the variance to wonder due to time-varying
estimation skill. This is an issue we need to discuss at some length. For
example, the 1x1 CRU grids of interpolated climate fields "relax" towards
climatology as the interpolation skill decreases. So do Kaplan's
interpolated SST and SLP grids. And, Mike Mann's 1000-year NH temperature
estimates do the same thing. That is honest in the sense that it truly
reflects the loss of estimation skill back in time. On the other hand, it
makes any use of the data very difficult because such trends in variance
will clearly affect any kind of natural variability analyses. As it is now,
we routinely "correct" the variance of tree-ring chronologies for changing
sample size, which is not that much different from what I did to the
extended PDSI recons. But I am still not very satisfied with any of this.
Any thoughts on this matter would be gratefully received.
I will put the gridded instrumental PDSI on my ftp site soon. It is not
handy just now.